Qwen3 TTS reaches Together AI with first audio in 97 ms and Apache 2.0 license
Together AI has hosted Qwen3 TTS, Alibaba's open voice synthesis model under Apache 2.0 license, with promise of first audio in only 97 ms.

Key takeaways
- O Together AI passou a hospedar o Qwen3 TTS, modelo de síntese de voz da Alibaba.
- O Qwen3 TTS promete primeiro áudio em apenas 97 ms.
- O modelo é distribuído sob licença Apache 2.0, permitindo uso comercial aberto.
- O artigo também aborda o Open VoiceFinder no ecossistema do Qwen3 TTS.
Together AI now offers Qwen3 TTS, an open-voice synthesis model released by Alibaba under Apache 2.0 license, with weights available on Hugging Face. The platform promises first audio in only 97 ms and released Open VoiceFinder, a tool to preview and compare voices before choosing which one goes to production.
Qwen3 TTS in Together AI: what the platform offers
The hosted inference platform included in its catalogue a text-to-talk model with open weights. In practice, this means that the developer can access Qwen3 TTS via API without providing his own infrastructure, or lowering the weights on Hugging Face and running the model wherever he wants — including on-premise.
The Apache 2.0 license is the central point here: it is permissive, allows commercial use without royalties and does not impose the typical license restrictions of models with closed weights. The announcement reinforces the expansion of open multimodal models in the hosted inference market, a movement that was already gaining traction with open weight alternatives for customization, such as Microsoft 'Licensed Weights'.
Qwen3 TTS: why 97 ms latency change experience
In voice agents, the metric that weighs the most is not the final audio quality, but the time to the first sound fragment. It is he who defines whether a conversation seems natural or if the user waits in silence. The number released by Together AI — 97 ms — belongs to the supplier himself and places Qwen3 TTS in the real-time conversational application range.
The level follows a broader race for minimum latency in inference, a theme that already appears in movements such as partnership between Cerebras and Oracle to accelerate AI responses. It is worth noting that the ad does not detail the test conditions or the hardware used in the measurement.
Open VoiceFinder: what is it for in the Qwen3 TTS ecosystem
Together AI released a voice preview tool together. Instead of integrating the model, generating samples and comparing files manually, the developer tests and compares timbres within the platform itself before deciding which configuration to use.
This shortens the cycle of iteration in products that depend on voice — automated service, audiobooks, reading screens for accessibility, content narration. The choice of voice is no longer an isolated technical test and becomes a product design step.
How Qwen3 TTS stands between open voice models
Open voice models compete directly with proprietary speech synthesis APIs, which charge for use and do not allow exporting the model. The advantage of Qwen3 TTS is the combination of lowerable weights and the convenience of a hosted API: the team can start fast and migrate to its own infrastructure later, if needed.
The announcement, however, focuses on latency and the voice tool. Price per character or per second of audio, number of available voices and list of supported langauges are not the focus of the text released — information that usually weighs on the decision of who will put the model into production.
Qwen3 TTS and Apache 2.0 license: why does it matter in Brazil
Open-voice synthesis reduces dependence on foreign suppliers with a fixed dollar price, a recurring problem for Brazilian teams operating on a tight budget. With the weights in hand, it is possible to host the model locally and keep the audios of the users within the company's own infrastructure.
This last point has regulatory weight: voice recordings can be treated as sensitive personal data under the LGPD, and processing everything on own servers simplifies conversation with the legal. For accessibility and education products, a permissive TTS also reduces the cost per hour of narrated content.
Qwen3 TTS FAQ
What is Qwen3 TTS?
It is a voice synthesis model (text for speech) released by Alibaba, with open weights under Apache 2.0 license and available on Hugging Face. It can now be used via API in Together AI.
What does the latency of 97 ms mean?
It is the time until the first audio is generated after sending the text, according to Together AI. This interval is decisive for voice agents to sound natural in real-time conversations.
Do I have to pay for the model to use the pesos?
Nope. The Apache 2.0 license allows you to download and use the weights commercially without royalty. What you pay is the hosted inference when you choose to use the Together AI API instead of running the template on your own.
Source: Together AI.
Source: www.together.ai

Written by
Lucas MontarroiosSou Lucas Montarroios e dediquei os últimos 15 anos à linha de frente de operações de telecom e data centers. Minha carreira sempre foi pautada por um foco implacável: transformar tecnologia e cenários críticos em oportunidades reais de negócio. No novidades.ia.br, trago essa visão executiva para o universo da IA. Especialista em produtos, mercado e ferramentas práticas de IA. Minha missão aqui é filtrar o ruído do mercado, analisando benchmarks, estratégias de grandes empresas e ferramentas práticas.


